Speaking AI

Two half-days and AI stops being a matter for specialists

Not a lecture. A training session where you handle a real document-based AI, where every technical term translates into a decision issue and where you leave with your first steps in black and white.

the outcome

The training memo

A workbook each participant receives in digital form and picks up again on Monday morning.

digital made clear — TRAINING MEMO Speaking AI at [your company] Training for decision-makers Two half-days, one memo PARTICIPANT COPY [date] · version 1.0 · Pierre-Yves Niederhauser
[your company] Speaking AI — 01 · A BRIEF HISTORY Older than it looks 1950 The Turing test can a machine fool us? 1956 The name is coined the term is born at Dartmouth 1997 Deep Blue beats Kasparov raw computing power wins at chess 2012 Deep learning machines learn to recognize images 2017 The transformer the architecture behind today's models 2022 ChatGPT AI enters everyday language What burst out in 2022 started back in 1950. PARTICIPANT COPY page 2
[your company] Speaking AI — 02 · GLOSSARY Ten key terms LLM The engine: text in, text out. Itknows nothing about you. token The unit text is cut into. It setswhat you pay. vector Text turned to numbers: the machinecompares ideas, not words. contextwindow What the model holds in mind at once.Beyond that, it forgets: see p. 4. hallucination A wrong answer, said with aplomb.Can be managed, not eliminated. RAG The engine wired to your documents.You will try it on p. 5. The session handout details all ten key terms. PARTICIPANT COPY page 3
[your company] Speaking AI — 03 · RAG Your documents become a memory you can query 1 · YOUR DOCUMENTS CONTRACTS MINUTES PROPOSALS SPLIT INTO VECTORS · P. 3 2 · THE MEMORY A MEMORY YOU CAN QUERY YOUR QUESTION 3 · THE ANSWER A written answer, followed by its sources: ▸ Master agreement 2024 · p. 12 ▸ Board minutes, 14 March · p. 2 RAG does not invent: it retrieves and cites. PARTICIPANT COPY page 4
[your company] Speaking AI — 04 · THE DEMONSTRATION Mécanix in your hands Mécanix SA is a fictional company DEMONSTRATION, THEN HANDS-ON YOU What does our maintenance contractprovide in case of breakdown? THE ANSWER Corrective intervention within 48 workinghours, essential spare parts included.The clause is identical in both activecontracts held by Mécanix SA. A reportfollows every intervention. SOURCES CITED ▸ maintenance_Mazak_2025.pdf · Wenk AG ▸ maintenance_Zeiss_2025.pdf · Zeiss Industrial IN SESSION You ask your own questions, on your own topics. PARTICIPANT COPY page 5
[your company] Speaking AI — 05 · THE ARCHITECTURES Three logics, complementary RAG EXTENDED CONTEXT DATA LAKE Do answers citetheir sources? YES PARTLY NO Ready in a fewweeks? YES YES NO Entry costunder control? YES BY VOLUME HIGH Content stays freshwithout effort? CONTINUOUS EVERY TIME ONCE BUILT ADVANTAGE CONDITIONAL LIMITATION PARTICIPANT COPY page 6
[your company] Speaking AI — 06 · THE LEGAL FRAME What the law asks of you EU AI ACT · SWISS FADP UNACCEPTABLE PROHIBITED HIGH RISK REGULATED LIMITED RISK TRANSPARENCY MINIMAL RISK FREE Most SME use cases sit in the two lower tiers. BEFORE YOU SIGN Where is the data stored? Who can read it?Is it used to train the model? PARTICIPANT COPY page 7
[your company] Speaking AI — 07 · MONDAY MORNING What you will do Monday WRITTEN AT THE END OF THE SESSION · NOT AFTER ACTION Draft a first version of ourAI usage charter WHO management WHEN week 26 ACTION List the AI tools already in useacross the company WHO head of IT WHEN by end of June ACTION WHO WHEN This is the only page I do not write. PARTICIPANT COPY page 8
01
A brief history From Turing to ChatGPT: AI was not born yesterday
02
Glossary LLM, token, vector: translated into your words
03
RAG Your documents become a memory you can query
04
The demonstration An AI wired to real documents, in your hands
05
The architectures Three logics, their advantages and their limits
06
The legal frame What the EU AI Act and the Swiss FADP ask of you
07
Monday morning Your next steps, written and dated